Agentic, Autonomous Design Evolution of Machine-Learned Interatomic Potentials

Recent AI systems can autonomously execute scientific workflows and empirical software tasks. We investigate whether an autonomous agent can use evaluation feedback to iteratively develop a scientific model. To address this question, we developed AutoResearch-MLIP, a research protocol in which the agent follows predefined instructions to propose, implement, and evaluate MLIP designs, with outcomes carried forward to guide subsequent iterations. Starting from a simple pair-distance MLIP, each generation selects a parent model by balancing exploration and exploitation, while structured design records guide subsequent proposals. Across 40 generations, the agent explores MLIP designs spanning local message passing, radial representations, scalar-vector states, body-order terms, residual readouts, long-range-inspired terms, and periodic image-offset graph construction. The final G40 model is competitive with established MLIP baselines in post hoc benchmarks. These results show that, within predefined rules, an autonomous agent can use evaluation feedback to iteratively improve MLIP implementations.